AS LOCATION OF YOUTUBE SERVERS

• YouTube video servers are located in an Autonomous

System (AS)
• We employ the ‘whois’ tool to map the server IP
address to the corresponding AS.
• Note: Since 2009 Google has migrated most content
from the YouTube original infrastructure (that was
based on third party CDNs) to its own CDN
SERVER GEOLOCATION
• Techniques used to identify the geographical location of
the YouTube servers. First is how IP addresses of
YouTube servers may be mapped to the appropriate
YouTube data centers. The goal is to later use this
information to analyze the video server selection policies.

1. Limitations of IP-to-location databases

2. Measurement based geolocation mechanism
3. Geolocation Results
1. Limitations of IP-to-location
databases
• One common way to find the geographical location of an
IP address is to rely on public databases. While such
databases are fairly accurate for IPs belonging to
commercial ISPs, they are known to be inaccurate for
geolocation of internal IPs of large corporate networks.

• A recent approach was taken where the location of the

server is obtained directly from the server name, i.e using
DNS lookup. DNS names often indicate location.
However, this approach is not applicable to the new
YouTube infrastructure, where DNS reverse lookup is not
allowed.
2. Measurement based geolocation
mechanism
• CBG (Delay-Based Geolocation) algorithm is a well-
known geolocation algorithm that is based on simple
triangulation.
• . A set of landmarks is used to measure the RTT to a
target. A simple linear function is then used to estimate
the physical distance between each landmark and the
target. This distance will become the radius of a circle
around the landmark where the target must be located.
The intersection among all circles is the area in which the
target can be located.
• CBG is more adequate.
3. Geolocation Results
• In particular, servers are grouped into the same data
center if they are located in the same city according to
CBG. We note that all servers with IP addresses in the
same /24 subnet are always aggregated to the same data
center using CBG approach.
EVALUATING YOUTUBE’S SERVER SELECTION
ALGORITHM

• How user video requests are mapped to YouTube data

centers. We are interested in exploring the various factors
that can influence the decision, such as user proximity,
server load, and popularity of content.

A. Video flows and sessions

B. Server selection strategy
C. Mechanisms resulting in accesses to non-preferred
data centers
A. Video flows and sessions
• In the normal scenario, each YouTube video request
corresponds to a HTTP message exchanged between the
Flash Plugin and a content server. If the request
succeeds, then the content server starts to deliver the
video inside the open connection. It is possible however
that the server may not serve the content. In such a case,
it would simply redirect the user to another content server
and close the connection. Other possible responses from
the server is change of video resolution.
A. Video flows and sessions
• Video flows - an isolated video flow not preceded by
other control flows is an indication that the request was
directly served by the contacted server.
• Control flows – short connections with signaling
messages (redirection) and precedes video flows. Control
flows are identified
B. Server selection strategy
• The Data center that provides the more traffic is preferred
• The Data center with the smallest RTT is preferred.
• The others are labeled as non preferred.
• In most cases the data center with the most traffic is the
data center with the smallest RTT.

• However, In some cases, Data centers with high RTT are

preferred. RTT is not the only criteria.
C. Mechanisms resulting in accesses
to non-preferred data centers
• A non-negligible fraction of video flows are downloaded
from non-preferred data centers.

Possible causes:
1. DNS mechanisms direct a request to the non-preferred
data center. (mapped by a single video flow or begin
with a control flow)
2. Redirected to another data center by the preferred data
center server (begins with a control flow to a preferred
data center)